Performance improvement for 2-qubit GST about pygsti

[rileyjmurray]

Hi, @eendebakpt! Thanks for doing this digging and for your suggestion. We're swamped right now and expect to be swamped until March Meeting is over, so we can't promise immediate progress on this. That said, we discussed this issue in today's pyGSTi developer meeting and we came to the following conclusions.

We're generally supportive of simple changes that buy us 10% -- 15% savings in expensive routines. In this context we'd want to check a few things before implementing this fix:

• Do cache accesses work correctly when running with MPI?
• Does caching negatively impact MapForwardSimulator? (If it does we could just isolate the change to MatrixForwardSimulator.)
• Is there a deeper inefficiency here that we can resolve even more thoroughly than the caching you've demonstrated?

Of course, the last point is vague and tricky. It really just means "some subset of {Stefan, Corey, Erik} want to look at code profiling results to see if anything else jumps out to them."

With that in mind, can you send us the profiling data you generated? Whatever mechanism works for you is fine with us.